Set up your AI assistant once so you stop re-explaining your job

A new conversation with an AI assistant knows almost nothing about your work. It doesn’t know that your director reads only the first paragraph, or that you write to suppliers in Spanish and to head office in English. So you explain it all again, or you take an answer written for nobody in particular and fix it by hand.

Most assistants can keep a short note about you and read it before every conversation, and most let you set up projects: separate spaces with their own instructions and files for one kind of work. The assistant can even interview you to write the note.

By the end of this, in about an hour, you’ll have that note saved and tested against a chat without it, and this week’s run of one recurring job done inside its own project. The steps work in ChatGPT, Claude, Gemini and Copilot; the menus for each are further down.

What you’ll be able to do

• Save a note of about 250 words on your work that your assistant reads before every conversation.
• Show what it changes: one weekly request answered before and after, with the edits each answer needed counted.
• Run one recurring job, such as a weekly report or customer replies, inside a project with its own instructions and files.
• Clear out of the assistant’s memory whatever is out of date.

What you need

• The chat assistant you already use. A free plan is enough to start; the menus below note where a feature depends on your plan or type of account.
• 45 to 60 minutes, and a request you make every week, such as a reply to a customer or an update for your manager.
• Standing instructions travel with every conversation, so describe your work in general terms and keep client names and confidential details out of the note. Before real files go into a project, Before you upload a client document lists what to check.

Steps

1. Run the “before” test. Before you change any setting, open a new conversation, type your weekly request the way you normally would, and paste the request and the answer into a document.
2. Fill in the note. Copy the template below and fill in each line as if you were briefing a capable new colleague. Spend the most time on who reads your work and what you always end up correcting in drafts. A filled-in example follows the templates.
3. Let the assistant interview you. In a new conversation, paste the interview prompt below with your draft at the end. The assistant asks up to ten questions, one at a time, about whatever your note leaves vague, then rewrites it. Answer in detail, by voice if you like; Why your AI drafts sound generic has the longer version of this technique.
4. Trim it. Aim for about 1,500 characters, roughly 250 words. Check the count in your word processor; assistants are bad at counting. Delete every line that wouldn’t change an answer, starting with “be professional”, and write rules you could check instead, such as “emails under 120 words”.
5. Save it in your assistant’s standing instructions, using the menus below. If your tool has nowhere to put it, or the box won’t take the whole note, keep the note in a document and paste it at the start of any conversation where it matters.
6. Run the “after” test. Open a new conversation and send the same request, word for word. Count the edits each answer needs before you’d send it, and write both counts at the top of your document: that’s your own measure of what the note is worth. Where the new answer still misses, add a line to the note and test again. If nothing changed at all, check that the instructions are switched on.
7. Set up a project for one weekly job, such as a status report or replies to customers. Create the project, paste in instructions for that job only (there’s a made-up example below), and add two or three reference files: your template, a past example you liked, a list of names and terms. Repeat any line from your note that matters for the job, since some assistants put project instructions ahead of your general ones. Then do this week’s run of the job inside it.
8. Check the memory. Most assistants also remember details from past conversations, and the interview just told yours a lot about your work. Open the memory settings, read what’s stored, and delete anything out of date, such as an old job title or last year’s report format.

You’re done when your document holds both answers with their edit counts and this week’s run of the job has come out of its project.

Three templates to copy

The “about my work” note

ABOUT MY WORK (updated [month and year])

My role: [job title] at [kind of organization, no names]. [Who reports to me, who I report to.]

What I use you for most: [two to four recurring tasks]

Who reads my work: [the usual readers and what they look for]

How I write: [tone, sentence length, what comes first]

Formats: [length and layout for each kind of document]

Avoid: [words and habits you always delete]

Language: [languages, spelling, formal or informal address]

When you're unsure: ask me one question before answering. Never invent figures, names or dates; write [CHECK] where one is needed.

The interview prompt

Below is a draft of the standing instructions you'll read before every conversation with me. Interview me to improve it.

- Ask up to ten questions, one at a time, and wait for each answer.
- Focus on what would most change your answers: who reads my work, what a good result looks like, what I usually correct in your drafts, what to do when you're unsure.
- If an answer is vague ("professional", "clear"), ask what it means in practice.
- Don't ask for names of clients or colleagues, or for confidential details.

Then rewrite the note under the same headings, as plain instructions addressed to you, in under 250 words, keeping my wording where you can. List anything you left out.

MY DRAFT:
[paste your note here]

Project instructions for a weekly report (made-up example)

WEEKLY STATUS REPORT
(Made-up example. Replace every line with your own.)

Every Friday I paste my work log for the week. Turn it into the status report for the head of communications.

Files: report-template.txt is the format to follow. example-report.txt is a past report; match its length and tone, never its content. names-and-terms.txt lists project names and who's who.

Rules:
- Use only what is in this week's log. Add no work, causes, dates or plans.
- Blocked items and decisions needed go first, stated plainly.
- Copy every figure exactly. Don't round or calculate.
- Under 250 words. Afterwards, list any log lines you left out and why.

The log and the template come from A running log that turns into your weekly status report. For a customer replies project, the one-page style card makes a good reference file.

A filled-in note (made up)

For example, here is the note of a fictional customer service team lead at an online furniture store, after the interview and the trim: about 1,100 characters.

ABOUT MY WORK (updated October 2026)

My role: customer service team lead at an online furniture store. Eight agents report to me; I report to the operations manager.

What I use you for most: replies to complaints my agents pass up to me, the weekly team update, training notes for new agents.

Who reads my work: customers who are already annoyed and reading on their phones; the operations manager, who reads the figures and the first line of everything else.

How I write: warm and brief. The answer or the next step first, then the reason. Apologize once, in one sentence.

Formats: customer replies under 120 words, ending with one next step and a date; the team update as bullets under Done, Problems and Next week.

Avoid: "we apologize for any inconvenience", "rest assured", exclamation marks, promising refunds or delivery dates I haven't confirmed.

Language: American English. Spanish for customers who write in Spanish, using usted.

When you're unsure: ask me one question before answering. Never invent order numbers, dates or policies; write [CHECK] where one is needed.

“Warm” on its own would leave the assistant guessing. “Apologize once, in one sentence” is a rule you can check in the after test, and so is every limit in the Formats line.

Where to find it in each assistant

Menus checked in October 2026. They change often, so if a name doesn’t match, look for the nearest equivalent in your settings.

ChatGPT

• Instructions: Settings > Personalization. Paste the note into Custom instructions and check that customization is switched on.
• Projects: New project in the sidebar. The project’s settings hold its instructions, and you can add files to it.
• Memory: Settings > Personalization, under Memory.

Claude

• Instructions: click your initials at the bottom left, then Settings, and paste the note under Instructions for Claude.
• Projects: Projects > + New Project, then Set project instructions; add files with the + button on the project page. Free accounts can have up to five projects. If you don’t see the instructions button, add the instructions to the project as a short text file.
• Memory: Settings > Memory lists what Claude remembers under Topics; open one to edit or delete it. To keep a one-off chat out of memory, open the + menu in a new chat and switch Memory off before the first message.

Gemini

• Instructions: on gemini.google.com, Settings & help > Personal Intelligence > Instructions for Gemini > Add.
• Projects: notebooks. Open the sidebar and, under Notebooks, click New notebook; add files with Add sources, and write the instructions under More > Notebook settings. If you don’t see notebooks yet, a Gem keeps its own instructions and files (Gems > New Gem, then Knowledge > Add files), but from November 2026 Google turns the Gems of personal accounts into skills.
• Memory: Settings & help > Personal Intelligence, where the Memory switch lets Gemini learn from your past chats, if your account has it (it needs a personal Google account with Keep Activity switched on). To stop Gemini from bringing something up, delete the chats that mention it from Gemini Apps Activity.

Copilot

• Instructions: in Copilot at work, signed in with your work account, Settings and more (…) > Chat settings > Personalization > Custom instructions. In the free Copilot app, look for personalization in the settings; the menus differ between versions.
• Projects: depending on your organization’s plan, Copilot at work may offer notebooks or agents that keep their own instructions. Without either, keep the job’s instructions in a document and paste them at the start of each conversation.
• Memory: at work, Settings and more (…) > Chat settings > Personalization > Manage saved memories. Your organization can switch personalization off, and then these settings show as off.

What can go wrong and how to check

The instructions fade in long conversations

Assistants follow standing instructions loosely, and often less closely as a conversation grows. Start a new conversation for each task, and repeat any rule that matters in the request itself.

The note tries to cover everything

The longer the note, the more each rule competes with the others, and a rule like “always use bullet points” turns a thank-you email into a list. Keep only the rules that suit almost everything you write, and move the rest into the project for that job.

The note and the memory go out of date

A new role or a new report format, and the note doesn’t know; that’s why the template starts with a date. Set a monthly reminder to reread it and rerun the after test. Deleting a line from the note doesn’t delete what the assistant remembered from the conversation where you wrote it, so repeat step 8 each time.

Where to go next

Your context is saved. The requests you repeat every week are the other half, and these pieces build on today’s setup:

• Keep your best prompts in one place gives those requests a home next to the note, two clicks from a blank chat.
• After your next meeting, give the setup a real test with Take meeting notes in three letters and send the follow-up within the hour, and count the edits the draft needs, as you did in step 6.
• If you’re not sure which weekly job deserves the next project, How to find your most repetitive work ranks your recurring tasks, starting from last week’s calendar.